[clang] Patch series to reapply #118734 and substantially improve it (PR #120534)

[Reid Kleckner via cfe-commits]

https://github.com/rnk approved this pull request.

Thanks for optimizing the builtins! I feel like builtins have grown significantly since adding RISC-V and ARM MVE intrinsics, and few people until now have stopped to re-evaluate how we represent these things.

While I was watching these clearly generated Intrinsic files land in Clang, I was thinking, these are clearly generated from a data model using a script. Wouldn't it be more efficient to encode the underlying data model for these operations using a more normalized representation? I proposed this at one point, and someone from Intel assured me that these long repetitive instructions and intrinsics are hand-maintained. It must be true at some level. Regardless, this change seems like a positive step towards renormalizing the otherwise denormalized builtins databse.

There's a lot of code and I'm sure there are more nits I could pick out, but I think it makes sense to reland this with the more sharded approach and get some feedback in tree.

https://github.com/llvm/llvm-project/pull/120534