Optimizing redundant calls to _tls_get_addr()

Optimizing redundant calls to _tls_get_addr()

[Roy Jacobson]

Hi,

I've recently encountered a performance issue with gcc and accessing
thread-local storage. I've posted a question on SO:
https://stackoverflow.com/questions/67894898/why-doesnt-gcc-eliminate-successive-calls-to-tls-get-addr
with the details of the problem, which has unfortunately not been answered
yet.

In short, GCC will call _tls_get_addr() on each access to a tls variable
inside a function, even if the variable is local. I would expect that from
the second access onwards, the result from the first access can be cached
and used instead. This is also what clang will do with "-O3".

A person on SO has led me to this discussion on the gcc mailing list in
2012: https://gcc.gnu.org/legacy-ml/gcc/2012-10/msg00024.html

I wonder what it would take to implement this optimization given the
current state of GCC? Are there maybe any thread safety reasons I miss why
it can't be done?

Thanks,
Roy Jacobson