[LLVMdev] Multi-threading and mutexes in LLVM

[Chandler Carruth]

On Fri, Jun 6, 2014 at 10:57 PM, Kostya Serebryany <kcc at google.com> wrote:

> As for the deadlocks, indeed it is possible to add deadlock detection
> directly to std::mutex and std::spinlock code.
> It may even end up being more efficient than a standalone deadlock
> detector --
> but only if we can add an extra word to the mutex/spinlock object.
> The deadlock detector's overhead comes primarily from two tasks:
>   1. get the metadata for the lock in question.
>   2. query the lock-acquisition-order graph to see if there is a loop.
>
> If the lock-acquisition-order graph is sparse (99% of cases we've seen),
> then the task #1 may constitute  up to 50% of the overhead.
> If we can add a word to std::mutex/std::spinlock data structures then the
> task #1 becomes trivial.
>

I don't see any reason not to reserve a word in the mutex so that in (an
ABI-compatible) debug build the mutex can support deadlock detection. Some
people are super concerned about having an extra word in a mutex, but I'm
not at all.

For libc++, it would probably need to be behind an ABI-breaking macro on
Mac and FreeBSD, but we haven't committed to any ABI stability on Linux, so
we could probably enable it by default there, and get into build bots.

Maybe bring this up on the cfe-dev list to discuss with Marshall and other
folks interested in libc++?
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20140606/fa94e597/attachment.html>