[sylware]

The compiler does not know the inner layout of how the system TLS variables are stored (could be musl way, bionic way, glibc way, etc). It only knows that for system TLS variables it has to go thru __tls_get_addr(), stated by the sysv ABI. Indeed, you are not forced to use pthread TLS to cache the value, you can use your own thread allocated memory, or keep it around on the stack/regs. On x86_64, once you have resolved the address with __tls_get_addr() you can use the same address value for all threads, then use common to all thread storage.

I think we agree.