[llvm] [GVN] Speed up GVN by caching DT nodes (PR #112903)

[Lewis Crawford via llvm-commits]

LewisCrawford wrote:

The 15% measurement was from Oct 2021, and it looks like the change to improve BB->DTNode lookup was c2f92fa3ab4 from https://github.com/llvm/llvm-project/pull/101705 from Aug 2024.

It looks like using the block numbers there might mean this DT caching is no longer necessary. It still removes 1 layer of indirection, but might not be worth the extra overhead of storing the DTNode pointers on the nodes themselves if it's just saving a vector lookup.

https://github.com/llvm/llvm-project/pull/112903