It does not optimize well

I think a nested call to `FOO.with` should be cached and create `tlv_get_addr` call only once, but every call becomes `tlv_get_addr` call and it hurts the performance.

```rust
FOO.set(&1, || {
FOO.with(|foo| {
use_foo(foo);

FOO.with(|foo| {
use_foo(foo);

});

});
});
```

I think this code should use some sort of cache for the thread local.

This is the actual profiling result