[llvm] [Coroutines] Split retcon ramps before ThinLTO IPO (PR #227232)

[Karim Alweheshy via llvm-commits]

karim-alweheshy wrote:

Thanks; I had not incorporated this alloca approach when I pushed the early-split revision. I rechecked Coroutines.md: its returned-continuation section says non-unwind coro.end implicitly returns a null continuation and shows unreachable after coro.end, including the retcon-once result example. Is your suggestion to change that pre-split retcon IR contract so a ret of a load follows coro.end? Also, would the return-value alloca replace the existing caller-provided buffer operand of coro.id.retcon, or require an additional way to escape it? I want to understand how this would remain compatible with existing bitcode. The current draft instead materializes the returns by splitting retcon coroutines before IPO; I am happy to prototype the alloca representation if that is the preferred direction.

https://github.com/llvm/llvm-project/pull/227232