On 6/17/13 1:36 PM, TommiT wrote:
std.algorithm.copy adheres to the specification of the C++ standard
library function std::copy. That specification states that the source
and target ranges may not overlap. Yet, all the major C++ standard
libraries (libc++, libstdc++, Dinkum C++ Library) implement std::copy so
that it becomes a memmove if the source and target element types are the
same and the element type is std::is_trivially_copy_assignable. Thus,
the current C++ specification seems to be too strict.
The current std.algorithm.copy implementation doesn't get optimized into
a memmove when it would be safe to do so. It really should, because it's
the fastest way to get the job done, and it allows the ranges to
overlap. Also, the documentation should be changed to notify of this
relaxation.
There's some benchmarks of memmove vs other methods:
http://nadeausoftware.com/articles/2012/05/c_c_tip_how_copy_memory_quickly
Bugzilla. FWIW I did notice Duff can sometimes beat memmove, but that
was some 10 years ago.
Andrei
