[llvm] [ThinLTO] optimize propagateAttributes performance (PR #132917)

[Zhaoxuan Jiang via llvm-commits]

nocchijiang wrote:

I tested the 2 ideas from the review comments. Unfortunately neither of them make noticeable improvements and I believe `DenseSet` overhead has to be avoided with a reasonable cost. Moving `MarkedNonReadWriteOnly` to `ValueInfo` and redesigning the inline flags to fit RO/WO/MarkedNonRWO into a 4-case enum results in a noticeable performance hit - now it takes 1 minute to complete `propagateAttributes()`. It appears that we must sacrifice some memory in order to enhance the performance. Thoughts?

https://github.com/llvm/llvm-project/pull/132917