[llvm] [CodeGen] Eliminate redundant spills at control-flow joins (PR #226852)

[Jeremy Morse via llvm-commits]

jmorse wrote:

As I'm a debug-info person I'm not in a position to review the register allocator changes; I can look at the LiveDebugVariables part, however I also don't really understand what's changing here. I'm unfamiliar with AMDGPU so don't really understand the tests added, and I also don't understand the notation in the body of the PR. Would you be able to produce a small x86 example of what's changing here to make it a bit easier. Understanding a simple case will help me understand the more complicated cases.

More generally, this sounds a reasonably serious change to part of the register allocator. It would probably be helpful to have an explanation of how often the affected code occurs, the performance impact, how much this patch improves it, and why this isn't solved by some other part of the register allocator.

https://github.com/llvm/llvm-project/pull/226852