Re: [Libunwind-devel] Updated fast trace patch with initial performance results

[Lassi Tuura]

Hi,

> Is there a particular reason you didn't use __thread?

Ah forgot one important detail. We do need to arrange a call to 
trace_cache_free() on thread exit. This is easy with pthread_key_create(), but 
I understand not so easy for __thread.

Any thoughts how to free per-thread data with __thread? Any tricks to get the 
destructor call effect without risking the calloc() call?

In our application we can arrange a call to libunwind on thread exit to free 
any per-thread caches. We could also arrange a call on thread creation, to 
ensure any calloc() for TLS will happen in 'safe' call context. These are not 
very generic solutions though.

Regards,
Lassi