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

[Nikita Popov via llvm-commits]

nikic wrote:

Do you see this 15% improvement against current LLVM main or some older version? The dominator tree has recently been improved to make use of basic block numbers, which makes the BB -> DT lookup significantly cheaper, so I find it quite surprising that you're seeing this level of improvement.

This causes a crash in the stage 2 build: https://llvm-compile-time-tracker.com/show_error.php?commit=d4557608b3a0c3b455c4d298b60a8c47a7de144f

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