[llvm] [SLP] Reland: More accurately cost RISCV scalar splats (PR #224766)

[Ryan Buchner via llvm-commits]

bababuck wrote:

> > > After this lands would you mind keeping an eye on LNT tomorrow and seeing if the performance is affected?
> > 
> > 
> > Looks like there are some regressions, I'll look into them: https://cc-perf.igalia.com/db_default/v4/nts/1117
> 
> Thanks, it seems like the povray regression was recoverered but namd and x264 are still regressed from the looks of things: https://cc-perf.igalia.com/db_default/v4/nts/1123?compare_to=1105

I don't see anything that appears actionable. Some of the regressions are from #221598 as well. From what I can tell, nothing seems to be misbehaving though with respect to the changes made. I did notice the SLP doesn't charge for non-splat constant vectors which seems wrong from SLP, will fix that.

https://github.com/llvm/llvm-project/pull/224766