[llvm] [WIP][SLP] Allow external uses to exist as both an extract and a rematerialization during the pass (PR #211680)

[Ryan Buchner via llvm-commits]

bababuck wrote:

> After some thoughts, I think we again need to start with the NFC patch. We need a class, which will represent instruction and it should be used to get the opcode, type, operands, etc. I mean, instead of asking actual instruction, we need a class that will hold this information. It will be useful for reductions and this approach too. And we need to start with this NFC aggregate and update of the instructions uses, and then extend this class to support dual representation

Is the idea that we want a lightweight SLP-IR that we can work with/modify and then only deal with the LLVM-IR when outputting vectorized instructions?

Would we use this new class everywhere in SLP or just select places?

For reductions, are you referring to the `zext-shl-and` case we just looked at or something else?

https://github.com/llvm/llvm-project/pull/211680