Re: [Qemu-devel] [PATCH 5/7] tcg-i386: Implement deposit operation.

[Alexander Graf <agraf@suse.de>]

Richard Henderson wrote:
> On 01/26/2011 11:01 AM, Alexander Graf wrote:
>   
>>> As far as I know, it does not happen at random.  Which seems to be 
>>> what you are suggesting.
>>>       
>> It happens on load/store and potentially helpers. The main difference
>> IIUC between globals and temps is that globals are kept in registers as
>> long as possible (read: until load/store or helper or tb end gets
>> emitted) while temporaries are not stored back to any memory, so they
>> are lost on load/store.
>>
>> So what you are suggesting is basically to use a different set of
>> globals for regs32 and to keep track of their usage throughout the TB,
>> so we can convert on demand. We can't use temporaries for that unless we
>> manually store them off on load/store/helper/tb end which means we'd
>> rewrite the globals treatment in target code :).
>>     
>
> No, what I'm suggesting is manually storing the reg32 temporaries back
> to their reg64 origins in the translator immediately before issuing the
> load/store/helper/tbend, at which point the generic TCG bits write back
> the reg64 globals to their env origin.
>   

Oh, I see :). That makes sense now. Thanks for the clarification :).

> Do you have a pointer to your s390x tree?
>   

Sure.

  git://repo.or.cz/qemu/agraf.git  s390

Please be aware of the fact that I'm currently reworking the whole CC
model, so if you start working on the register acceleration now there
will be conflicts for sure :). Do you have code to test it with?


Alex