RFR(M): 7197327: 40% regression on 8 b41 comp 8 b40 on specjvm2008.mpegaudio on oob

[Vladimir Kozlov]

On 1/22/13 8:21 AM, Roland Westrelin wrote:
>>> But 2- can bring more opportunities to do 1- again.
>>
>> Yes, but loop opts executed several times. Why you want to do everything during one iteration?
>
> It can happen without loops, with imbricated ifs.

You are right.

> Why would we want to have multiple loop opts passes in that case if we can do everything at once?

Changing controls of one set of expensive nodes changes the dominators 
information of whole graph. So, strictly speaking, you can't use this 
information for following sets of expensive nodes.

> What is the benefit of doing this in build_loop_early() rather than in process_expensive_nodes?

I think find best control could be viewed as independent optimization 
from finding common control. And build_loop_early() logically is the 
best place to do it. It will also help better place nodes during 
build_loop_later().

Vladimir

>
> Roland.
>