[llvm] [Instructions] cache computed shufflevector properties (PR #115536)

[Princeton Ferro via llvm-commits]

Prince781 wrote:

Hi @nikic,

We have some internal kernels that make heavy use of very large vectors (256, 512, 1024 elements). We find a 20-40% compile time reduction on these by caching shufflevector mask properties. I think this impact is hard to see with smaller sizes.

https://github.com/llvm/llvm-project/pull/115536