On Wed, Jun 3, 2009 at 8:50 AM, Thomas Klimpel <Thomas.Klimpel@synopsys.com>wrote:
I know that there are some geometry libraries currently under development that plan to be proposed for review sooner or later. At least one of them tries to address the robustness problems that occur in this context with a "novel" modification of an established technique. Should we tell the author that he has to publish these modifications first in a reputable journal before proposing his library for review? I don't think so. However, he will certainly have to convince the reviewers during the review that his implementation is reliable, which includes that the implemented algorithms are reliable.
I also know that this case is different from the one currently discussed, in that this "novel" modification is an implementation detail of the geometry library, while the "novel" algorithmic improvements are an important motivation for proposing the "sorting" library for review.
string_sort would work okay without any novel modification from American Flag Sort, but wouldn't be much faster than std::sort, due to the overhead of its max/min checking. I agree though, integer_sort and float_sort wouldn't be safe to use with any other published algorithm I'm aware of due to poor worst-case performance without the change to use std::sort, and increase of the fallback sort size (is that really so novel? Or just modernizing the algorithm?). The worst-case performance check is less essential once that step is taken, and does nothing on 32-bit integers with current default settings.