[llvm] [VPlan] Reject partial reductions with invalid costs in getScaledReds. (PR #180438)

[Florian Hahn via llvm-commits]

fhahn wrote:

> > Argh, really sorry I missed the connection to this!
> 
> Thanks, that's okay. Apologies from my end if I sounded like a bit of grump on the Monday morning raising this! :)
> 
> > I'm not planning any further refactoring, [...] One thing that is still missing w.r.t. partial reductions I think is that we probably should also move creation of expression recipes for them to directly `createPartialReductions`, to only convert the whole chain, if we also actually create combined expressions.
> 
> If you're not planning any further refactoring, then I'm going to clean up partial reductions a bit in the next 1-2 weeks. I was thinking of revisiting the cost interfaces and also see if parts of it can be stream-lined with the in-loop reductions. I've also got a WIP patch that is a step in the direction that you mention above. I'll share it shortly!

Sounds great! I think forming the VPExpressions together with transforming to partial reductions should address one of the open issues where we create partial reductions but then fail to create VPExpressions, which can end up worse.



https://github.com/llvm/llvm-project/pull/180438