Wilco (Wilco.Dijkstra@ntlworld.com) on 9/3/09 wrote:
---------------------------
>none (none@none.com) on 9/3/09 wrote:
>---------------------------
>>Wilco (Wilco.Dijkstra@ntlworld.com) on 9/3/09 wrote:
>>---------------------------
[...]
>>You must be kidding or only considering user level emulation.
>>And even for user level simulation, I would be surprised if
>>you could reach that level of efficiency. Don't forget that
>>even for user processes simulation you have to handle signals
>>and SMC. As a hint look at what DynamoRIO is able to achieve
>>without having to translate anything.
>
>I meant full CPU simulation - with user level you can get native performance using
>the simplest interpreter if you spend most of the time in the native OS and libraries...

So you can beat non-translating user binary inspection
tools and including MMU and hardware simulation. Extremely
impressive! Can I take a look at that software? :)

>>My very preliminary tests with QEMU running on Cortex-A8@500
>>show it can simulate nbench at about 486@50. Of course this
>>needs tuning, but I seriously doubt it can be sped up by 9 as
>>you imply :) BTW how much faster is 486 vs 386?
>
>QEMU isn't an optimizing JIT AFAIK, so it won't be able to get near static binary
>translation performance. A 486 is about 2-3 times as fast as a 386, so it's about
>a 100-150MHz 386. Still I think you're going to have to do a bit more tuning...
>Where do you lose all those cycles? Most x86 instructions translate into a few ARM
>instructions. You don't have to worry about simulating x86 flags most of the time for example.

QEMU codegen is indeed weak. But as I hinted in my previous
post, the cost isn't necessarily in running generated code.
Handling interrupts, taking care of potential SMC, indirect
branches, translated code cache capacity management and so
on have a big cost. If you don't handle all of that
correctly then your emulator is not transparent and will
fail to run some programs, which I consider as not
acceptable.

But there's definitely room for improvement. I just don't
believe a 3-4 speedup is achievable even for user mode code.