echelon


Move Now

Last week it was my great pleasure to attend the C++Now conference in Aspen. This is my trip report.

Due to snow the day before the conference and the remote location of Aspen, the conference got off to a rough start for most attendees. But almost everybody arrived in time. The conference got going with an interesting keynote by Gabriel Dos Reis and got better from there. Michael Caisse’s tutorial about the canonical class set the stage for what would become one of the central themes of the conference. The title of this report reflects this theme as well.

aspen-28better
Aspen, C++Now 2014, image by Chandler Carruth

Just what does it mean to move an object? In particular, what can you expect from the thing that you moved from? And what is this thing anyway? Eric Niebler had a well thought out opinion on this topic. He argued that a moved-from object should not break its invariants. This makes it a lot easier to reason about code. Not everyone agreed with this. Sean Parent for example argues, that a programmer can only expect to destroy and reassign a moved-from object.

Following this reasoning, he presented an unsafe move operation that does not change the moved-from object. This is unsafe because bad things will happen if such an object is destroyed and then the moved to object is used or destroyed. So the unsafe move requires the programmer to do something with the moved from object – usually moving another object there. Sean stated that this trick can improve performance by up to 15% when for example reverse is called on a large container. David Sankel responded to this proposal in his talk with a simple “No!” and presented a better solution. Instead of breaking the invariants of the container, he added another container that does not have the original invariants and thus allows unsafe operations. For a list, such a container would be called ListFragments. Fragments are extracted from the list, some operations not legal on lists are performed on the fragments, and the fragments are put back into the list. This way, no invariants are broken. And Sean Parent approves of this idea.

For me, the second theme of the conference was asynchrony. Christophe Henry, the author of the Boost Meta State Machine library presented his latest creation Boost Asynchronous. In an exciting talk he convinced his audience that most approaches to asynchrony are doomed, and that he thinks he has found an elegant solution. From the discussion that followed his talk it became clear that there are some similarities between his work and that of the HPX library. Another talk on that topic was about the libcppa project and its application in a high performance network security auditing database.

Followers of the gospel of functional programming could enjoy the excellent talks by David Sankel and Bartosz Milewski. Louis Dionne presented initial results from his work on MPL11, the metaprogramming library proposal for C++11 and onwards. Ábel Sinkovics presented his interactive metaprogramming shell based on Clang. These two meta-programming wizards delighted their audience with deep insights into the language and an excellent tool to explore and exploit the language.

My own talk about accelerator programming libraries on the last day of the conference went well. However I ended the talk with the feeling that something was missing. And indeed, I got some feedback that maybe some performance comparison of the different libraries I reviewed would be interesting. And that is what I am working on now.

All in all, C++Now was an educating and delightful experience. I hope to go back next year.



No Comments »

Leave a comment