During the review of JDK-8341630, we identified a ~20% performance regression in G1 Concurrent Mark From Roots times following the change. Further investigation revealed that the regression was caused by loss of inlining of G1CMTask::deal_with_reference

A similar regression was observed in the Valhalla repository, where preventing inlining of G1FullGCMarker::mark_and_push led to a comparable increase in Full-GC marking times.