Re: RISC-V: Folding memory for FP + constant case

[Jeff Law <jeffreyalaw@gmail.com>]

On 8/1/23 17:03, Vineet Gupta wrote:
> 
> 
> On 8/1/23 15:07, Philipp Tomsich wrote:
>> +Manolis Tsamis
>>
>> On Tue, 1 Aug 2023 at 23:56, Jeff Law via Gcc-patches
>> <gcc-patches@gcc.gnu.org> wrote:
>>>
>>>
>>> On 8/1/23 13:14, Vineet Gupta wrote:
>>>
>>>> I have some numbers for f-m-o v3 vs this. Attached here (vs. inline to
>>>> avoid the Thunderbird mangling the test formatting)
>>> Thanks.  Of particular importance is the leela change.  My recollection
>>> was that the f-m-o work also picked up that case.  But if my memory is
>>> faulty (always a possibility), then that shows a clear case where
>>> Jivan's work picks up a case not handled by Manolis's work.
>> f-m-o originally targeted (and benefited) the leela-case.  I wonder if
>> other optimizations/changes over the last year interfere with this and
>> what needs to be changed to accomodate this... looks like we need to
>> revisit against trunk.
>>
>> Philipp.
>>
>>> And on the other direction we can see that deepsjeng isn't helped by
>>> Jivan's work, but is helped by Manolis's new pass.
>>>
>>> I'd always hoped/expected we'd have cases where one patch clearly helped
>>> over the other.  While the .25% to .37% improvements for the three most
>>> impacted benchmarks doesn't move the needle much across the whole suite
>>> they do add up over time.
>>>
>>> Jeff
> 
> I took a quick look at Leela, the significant difference is from 
> additional insns with SP not getting propagated.
> 
> e.g.
> 
>     231b6:    mv    a4,sp
>     231b8:    sh2add    a5,a5,a4
> 
> vs.
> 
>     1e824:    sh2add    a5,a5,sp
> 
> There are 5 such instances which more or less make up for the delta.
ACK.  Jivan and I have seen similar things with some other work in this 
space.  What's weird is the bits of Manolis's work that went in a month 
or so ago are supposed to address this exact issue in the post-reload 
const/copy propagation pass.

Jeff