nikic wrote: Yeah, I'd like to see a reproducer for this as well. I suspect that this doesn't so much fix the problem as hide it for the specific test case. Two calls are more than enough for exponential inlining :) https://github.com/llvm/llvm-project/pull/189782