Re: [Qemu-devel] [RFC 38/48] translator: implement 2-pass translation

[Emilio G. Cota]

On Wed, Nov 28, 2018 at 12:50:23 +0000, Alex Bennée wrote:
> Emilio G. Cota <address@hidden> writes:
> > I just wrote some code to go over the list and add TB callbacks,
> > which go right before the first insn_start Op. The code is hack-ish
> > in that we first generate the TCG ops we need, which get added to
> > the end of the ops list, and then we go over those and move them
> > to where we want them to be (before insn_start, in this case).
> > But it works and it's less of a hack than doing the whole 2nd pass.
> 
> But we should be able to insert the ops directly in the right place.
> That is the whole point of being a list right ;-)

Right, it's just hard sometimes to know exactly where to insert.

> > Insn callbacks will be trivial to implement this way; memory
> > callbacks should be harder because there are several qemu_ld/st
> > opcodes, but it should be doable;
> 
> I was thinking about this last night. I wonder if we need to tag the
> memory tcg ops so we can find them afterwards during the insertion
> phase - but maybe the opcode itself provides enough information.

We should be able to extract the info from the memop argument,
I think.

> > last, memory instrumentation
> > of helpers might actually be easier than with the 2 passes, because here
> > we just have to look for a Call TCG op to know whether a guest
> > instruction uses helpers, and if it does we can wrap the call
> > with the helpers to generate the setting/resetting of
> > CPUState.plugin_mem_cbs.
> 
> So merging the two helper calls into one from the target code?

Actually we don't need helpers to set/reset CPUState.plugin_mem_cbs;
we do that in TCG directly. So here we could just add the "set"
code right after "insn_start", and the "reset" code at the very
end of the translation (right before tb_exit/goto_tb etc).
The "reset" might still be dead code, but that is unavoidable
because the helper might do a longjmp. But we can fix that by
resetting the variable when returning from the jump.

Thanks,

                E.