During the remark pause G1 traverses the string table, looking for dead oops. This is wasted effort because there cannot be any dead strings currently because:
- the initial mark uses the string table as roots (at the moment unconditionally, later with class unloading conditionally)
- SATB requires that any entries added later than initial mark and before remark must be kept live (and we are not treating them as weak roots anyway).
So there should be no gain looking for dead string table entries at all.
Removing this part of the remark saves ~13% of total remark pause times across a large range of FMW applications.