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

[Florian Hahn via llvm-commits]

fhahn wrote:

> Accepted with some nits.
> 
> On [#178919 (comment)](https://github.com/llvm/llvm-project/pull/178919#discussion_r2769196559) Graham pointed out this exact same issue and I wrote there that I would fix this in a separate PR, which I eventually spent a chunk of my Friday afternoon working on. Unfortunately I didn't have time to post it before I had to catch my train and was subsequently away all weekend. Because you left comments on that PR as well, I figured you would have seen the conversation and made the connection with the Github Issue that was reported. Did you miss that conversation or not realise these were the same issue? There's no point in me posting my changes now as they're largely identical, I just wish I would have spent my time on Friday differently.
> 
> As a general comment, the LV is a codebase that sees a lot of churn and a way for others to keep up is to help with fixing the odd issue. As a maintainer/area lead it may help to step back a bit sometimes and leave room for others to get involved. Otherwise, this is not really a collaborative project.
> 

Argh, really sorry I missed the connection to this!

> With that being said, are you planning any other refactoring or improvements for partial reductions? It would be good to avoid further duplicating work.

I'm not planning any further refactoring, the PR here also was not planned, but in respones to an issue filed on the weekend. I'll make sure to communicate better if any of that changes.

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.


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