G1 uses TLS storage to store the ever-changing current card table to modify.

The reason is that this generates the most compact code: inlining some address where the current value can be obtained has the disadvante of increases code size significantly, particularly on non-x64 where constructing a 64 bit integer takes many instructions. Even on x64, this is a 10 byte instruction.

This in turn at best has the same performance as using TLS (x64), worse on aarch64, even modern implementations.