C++ Logo

SG5

Advanced search

Subject: Next transactional memory meeting
From: Hans Boehm (boehm_at_[hidden])
Date: 2019-05-24 00:23:54


Please make sure that you are subscribed to sg5_at_lists.isocpp.org. In case
you missed it, we switched mailing lists!

We failed to get a quorum on April 29. The next meeting is currently
scheduled for Monday, May 27, Memorial Day in the US. I doubt we will get a
quorum then either.

Can we delay this meeting (only) by a week, and try for June 3? If you're a
regular attendee, please let me know if you can or cannot make a meeting on
the 3rd, at the usual time.

Hans

On Sun, Apr 28, 2019 at 9:00 PM Hans Boehm <boehm_at_[hidden]> wrote:

>
> We are meeting tomorrow April 29th at noon PDT, usual coordinates:
>
> Join Hangouts Meet
> meet.google.com/sbj-cvgh-vnd
> Join by phone
> ‪+1 208-925-0196‬ PIN: ‪255 542‬#
>
> The topic will continue to be the TM-lite proposal.
>
> Notes from March 25:
> -----------------------------
>
> Attendees: Jade Alglave, Hans Boehm, Michael Scott, Michael Spear, Victor
> Luchangco
>
> Discussed Victor's additions to
>
> M Spear: Synchronized instead?
>
> Victor: No, want atomic block.
>
> M Spear: Want to be able to use STL, e.g. map in transaction.
>
> Hans: trade-off, non constexpr makes it implementation defined
>
> Victor: Want nested transactions
>
> M Spear: Should allow proper software engineering in transactions.
>
> Victor: Should look at what's really allowed in constexpr.
>
> M Spear: Want everything visible in translation unit to be allowed.
>
> Hans : trade-off against portability. Constexpr growing.
>
> M Spear: Not for simplifying implementation. Single global lock always
> works.
>
> All: Aim for subset of atomic_noexcept.
>
> Do we need to catch transaction-unsafe code that's never executed?
>
> Hans: Previously decided on undefined behavior when executing
> transaction-unsafe code.
>
> M Scott: Is SGL with no checks conforming?
>
> Allow either detection or simple SGL?
>
> M Spear: Real challenge is getting users. This is simpler than existing
> implementations.
>
> M Scott: 2 possible implementations:
> HTM-centric plus diagnosis
> SGL with no diagnosis
>
> M Scott: Should warn, but not for atomics
>
> M Spear: Should prohibit volatile, atomics, inline assembly
>
> Victor: What about IO?
>
> M Scott: Don't want to require diagnosis for unexecuted code.
>
> Jade: Lock-free?
>
> Victor: Want to allow SGL for now
>
> Meet next time as scheduled, on April 29.
>
>



SG5 list run by herb.sutter at gmail.com

Older Archives on Google Groups