[llvm] [AMDGPU] SIFixSGPRCopies avoid placing V_READFIRSTLANE inside the loop. (PR #186400)

[Sameer Sahasrabuddhe via llvm-commits]

ssahasra wrote:

This pass is not something I am familiar with. But in general, since we are talking about a cost-based heuristic, is there an example where the RFL _will_ be inserted inside the loop? Can it be added as a test?

https://github.com/llvm/llvm-project/pull/186400