[llvm] [Inline] Accumulate the cost of the inlined function to the new call site (PR #111104)

[via llvm-commits]

DianQK wrote:

I think I've done something similar to the `function-inline-cost-multiplier`. However, I believe that idempotent makes sense here. Other approaches I can think of are:

- Lowering the inline threshold when the function itself contains plenty of instructions.
- Not applying the hot call site threshold when there is a loop inside the function at the call site.

In fact, I believe the main reason for the increased compile time is that we've inlined nested loops multiple times.

https://github.com/llvm/llvm-project/pull/111104