...
I have started experimenting with this. It provided a very small, barely significant (when lucky and there is very low noise in the benchmark) speed-up of about 0.5% on the old jsc-only benchmarks (Octane, Kraken, Sunspider). I would like to do more thorough benchmarking, including ARES/Speedometer before landing this change. I have also tested scaling the inliningCost by the tier reached by the callee (encouraging inlining of very hot functions, and discouraging inlining of very cold ones). So far it has not seemed to bring any measurable additional improvement.