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

[Arthur Eubanks via llvm-commits]

aeubanks wrote:

> Not sure I follow this bit. I don't think this makes the inliner non-idempotent by itself? The cost attribute is only adjusted if something actually gets inlined, so it shouldn't introduce any new non-idempotence.

not sure what I was thinking in 2024, maybe I was under the impression that there was some state only being kept in memory, rather than recorded into the IR

this seems like a pretty big inliner cost change, have you done any analysis on how much less inlining happens with this change?

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