Re: [PATCH 12/19] target/ppc/pmu_book3s_helper.c: enable PMC1 counter negative EBB

[Daniel Henrique Barboza <danielhb413@gmail.com>]

On 8/12/21 1:56 AM, Richard Henderson wrote:
> On 8/11/21 6:45 PM, Richard Henderson wrote:
>> On 8/11/21 5:39 PM, David Gibson wrote:
>>>> I mean, nothing is stopping us from calculating cycles using time, but in the
>>>> end we would do the same thing we're already doing today.
>>>
>>> Oh.. ok.  I had assumed that icount worked by instrumenting the
>>> generate TCG code to actually count instructions, rather than working
>>> off the time.
>>
>> David, you're right, icount instruments the generated tcg code.
>> You also have to add -icount to the command-line.
> 
> Oh, and btw, icount disables multi-threaded tcg, so you're going to be running that guest in round-robin mode.
> 
> Icount affects so many aspects of qemu that I really do not think it is the best option for a PMU.

Using icount in the PMU is my fallback plan. I spent some time trying to
count instructions using translationOps but never got it right. I got
up to a point where the tests were counting instructions for the first
time it was run in the guest, then nothing else counted in consecutive
runs.

I was able to figure out that it had to do with how the translation block
works. If a previously ran TB was found via lookup then the translationOps
callbacks I was using weren't being called. I know that I was missing a
piece of info there, but since I'm trying to deal with other aspects of the
PMU logic I fell back into using icount to get things of the ground.

All this said ....

> 
> If you want to count instructions retired, then just do that.  Stuff values into tcg_gen_insn_start so they're available for exception unwind, and otherwise decrement the counter at the end of a TB.

... I don't bother giving this a try. Can you clarify what do you mean
with "exception unwind"?

I tried something similar with tcg_gen_insn_start() (called via ppc_tr_insn_start()).
This this ops is called inside translator_loop(), and translator_loop() isn't
being ran if TB_lookup returns the TB, I was observing the behavior I
described above of a test counting instructions in its first run only.


> 
> If you really must interrupt exactly at 0 (and not simply at some point past underflow), then we can adjust the tb lookup logic to let you reduce tb->cflags.CF_COUNT_MASK in cpu_get_tb_cpu_state.

That would be good, but depending on the amount of work I would consider doing
this is a follow-up of this work. It's ok if the PMU overflows a bit instructions
for our current purposes ATM.


Thanks,


Daniel

> 
> 
> r~