Re: TSX lock elision for glibc v12

[Torvald Riegel <triegel at redhat dot com>]

On Fri, 2013-06-21 at 11:37 -0400, Rich Felker wrote:
> On Fri, Jun 21, 2013 at 05:00:29PM +0200, Torvald Riegel wrote:
> > Or should we just not try to support explicit initialization of DEFAULT
> > mutexes *via settype()* and just make them NORMAL mutexes by default?
> > This way, we wouldn't need to introduce a new public
> > PTHREAD_MUTEX_DEFAULT value, and could still disambiguate DEFAULT from
> > NORMAL when the initialization happens via static initializers or
> > default mutex attributes (ie, NULL parameter).
> 
> I'm fine with this. I don't think there's any sane reason to call
> pthread_mutexattr_settype when making default-type mutexes, and I
> doubt any real-world apps do it. If they do, it's their fault for
> being slow; they're performing unnecessary, bloated operations in any
> case and should be fixed.

Ok.

Attached is what I am currently testing.  This doesn't touch any of the
public mutex types, but disambiguates NORMAL from DEFAULT (modulo the
above).  All the mutex functions can then handle
PTHREAD_MUTEX_NORMAL_INT_NP and PTHREAD_MUTEX_DEFAULT_INT_NP
differently.  This should be sufficient as a base to enable elision for
the "90%", and we don't require further bits in the type.  IOW, it's
stage 1) in my three-stage proposal for how to commit those things we
can actually agree on now.

I hope to get back to this later today.

Attachment: patch
Description: Text document