[PATCH] D54889: Fiber support for thread sanitizer

[Dmitry Vyukov via Phabricator via llvm-commits]

dvyukov added a comment.

Thinking of how we can eliminate the indirection in cur_thread...

We could not even allocate a tid for fiber (tid space is expensive) if we record fiber switch events in the trace. Then all fibers running on a thread could reuse its tid, and we could restore fiber identity for memory accesses from trace. But it get pretty complicated at this point.

I think we sould split ThreadState into Thread part (which still can be called ThreadState) and Fiber part. ThreadState should contain at least fast_state, fast_synch_epoch, clock and maybe some of the int's and bool's (as they are easy to copy) and pointer to the current Fiber and to the physical thread Fiber. Fiber part will contain everything else. Fiber switch will switch Fiber pointer in ThreadState and copyout/copyin/update fast_state, fast_synch_epoch and clock.

This will eliminate the indirection on hot paths and will also open path to nested tasks (as Fiber can contain pointer to a nested Fiber).


Repository:
  rCRT Compiler Runtime

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D54889/new/

https://reviews.llvm.org/D54889