Re: [PATCH v2] rs6000: Add new pass for replacement of contiguous addresses vector load lxv with lxvp

[Ajit Agarwal <aagarwa1@linux.ibm.com>]

Hello Kewen:


On 05/12/23 7:13 pm, Ajit Agarwal wrote:
> Hello Kewen:
> 
> On 04/12/23 7:31 am, Kewen.Lin wrote:
>> Hi Ajit,
>>
>> on 2023/12/1 17:10, Ajit Agarwal wrote:
>>> Hello Kewen:
>>>
>>> On 24/11/23 3:01 pm, Kewen.Lin wrote:
>>>> Hi Ajit,
>>>>
>>>> Don't forget to CC David (CC-ed) :), some comments are inlined below.
>>>>
>>>> on 2023/10/8 03:04, Ajit Agarwal wrote:
>>>>> Hello All:
>>>>>
>>>>> This patch add new pass to replace contiguous addresses vector load lxv with mma instruction
>>>>> lxvp.
>>>>
>>>> IMHO the current binding lxvp (and lxvpx, stxvp{x,}) to MMA looks wrong, it's only
>>>> Power10 and VSX required, these instructions should perform well without MMA support.
>>>> So one patch to separate their support from MMA seems to go first.
>>>>
>>>
>>> I will make the changes for Power10 and VSX.
>>>
>>>>> This patch addresses one regressions failure in ARM architecture.
>>>>
>>>> Could you explain this?  I don't see any test case for this.
>>>
>>> I have submitted v1 of the patch and there were regressions failure for Linaro.
>>> I have fixed in version V2.
>>
>> OK, thanks for clarifying.  So some unexpected changes on generic code in v1
>> caused the failure exposed on arm.
>>
>>>
>>>  
>>>> Besides, it seems a bad idea to put this pass after reload? as register allocation
>>>> finishes, this pairing has to be restricted by the reg No. (I didn't see any
>>>> checking on the reg No. relationship for paring btw.)
>>>>
>>>
>>> Adding before reload pass deletes one of the lxv and replaced with lxvp. This
>>> fails in reload pass while freeing reg_eqivs as ira populates them and then
>>
>> I can't find reg_eqivs, I guessed you meant reg_equivs and moved this pass right before
>> pass_reload (between pass_ira and pass_reload)?  IMHO it's unexpected as those two passes
>> are closely correlated.  I was expecting to put it somewhere before ira.
> 
> Yes they are tied together and moving before reload will not work.
> 
>>
>>> vecload pass deletes some of insns and while freeing in reload pass as insn
>>> is already deleted in vecload pass reload pass segfaults.
>>>
>>> Moving vecload pass before ira will not make register pairs with lxvp and
>>> in ira and that will be a problem.
>>
>> Could you elaborate the obstacle for moving such pass before pass_ira?
>>
>> Basing on the status quo, the lxvp is bundled with OOmode, then I'd expect
>> we can generate OOmode move (load) and use the components with unspec (or
>> subreg with Peter's patch) to replace all the previous use places, it looks
>> doable to me.
> 
> Moving before ira passes, we delete the offset lxv and generate lxvp and replace all
> the uses, that I am doing. But the offset lxvp register generated by ira are not
> register pair and generate random register and hence we cannot generate lxvp.
> 
> For example one lxv is generated with register 32 and other pair is generated
> with register 45 by ira if we move it before ira passes.

It generates the following.
	lxvp %vs32,0(%r4)
        xvf32ger 0,%vs34,%vs32
        xvf32gerpp 0,%vs34,%vs45
        xxmfacc 0
        stxvp %vs2,0(%r3)
        stxvp %vs0,32(%r3)
        blr


Instead of vs33 ira generates vs45 if we move before pass_ira.

Thanks & Regards
Ajit

 
> Thanks & Regards
> Ajit
>>
> 
>>>
>>> Making after reload pass is the only solution I see as ira and reload pass
>>> makes register pairs and vecload pass will be easier with generation of
>>> lxvp.
>>>
>>> Please suggest.
>>>  
>>>> Looking forward to the comments from Segher/David/Peter/Mike etc.
>>
>> Still looking forward. :)
>>
>> BR,
>> Kewen