Re: [Qemu-arm] [Qemu-devel] [PATCH 5/7] target-arm: Add isread parameter to CPAccessFns

[Alex Bennée]

Peter Maydell <address@hidden> writes:

> On 5 February 2016 at 14:20, Alex Bennée <address@hidden> wrote:
>>
>> Peter Maydell <address@hidden> writes:
>>> -typedef CPAccessResult CPAccessFn(CPUARMState *env, const ARMCPRegInfo 
>>> *opaque);
>>> +typedef CPAccessResult CPAccessFn(CPUARMState *env,
>>> +                                  const ARMCPRegInfo *opaque,
>>> +                                  bool isread);
>>
>> I guess my only comment here is we've extended the call for every access
>> check with another parameter (and associated TCG activity) for something
>> only one handler currently cares about.
>>
>> Is there an argument for an rwaccessfn() that we use for just those
>> registers that care about the detail? I know system registers are hardly
>> a fast path priority but I'm concerned about knock on effects on
>> performance. Have you done any measurements?
>
> I haven't measured, no, but since there are only 3 arguments the
> third argument is going to be in a register on any host architecture
> we care about, which means the overhead is just going to be a single
> "load constant 0 or 1 into register before the call". I think that's
> going to be lost in the noise compared to actually having to make
> the function call at all, the work the function call does, and then
> the second function call later to do the read or write.

I was thinking of knock on effects on spilling other registers in the
TCG code. I guess this depends on how complex the code is around system
register access.

>
> thanks
> -- PMM


--
Alex Bennée