[llvm] [LoopVectorize] Don't discount instructions scalarized due to tail folding (PR #109289)

[John Brawn via llvm-commits]

john-brawn-arm wrote:

I'm giving up on this and instead going for the approach of disabling scalarization of predicated instructions when using optsize in #129265. The main problem here is that my assertion that the vector cost "is already the scalarized cost" isn't quite right - it's the scalarized cost assuming that none of the operands area scalarized and thus require an extract, and what computePredInstDiscount is doing is predicating instructions exactly so that we can get rid of the cost of the extract. I tried modifying this patch to account for this, by making the discount instructions that have already been scalarized be the cost of the extracts that will be removed, but didn't manage to make something that didn't make things worse in some cases (I think due to costs elsewhere being wrong, but figuring out what exactly was going wrong would take more time and effort than I'm willing to spend).

https://github.com/llvm/llvm-project/pull/109289