Is there any reason to keep the undefined behavior for shifts of negative numbers, instead of making it implementation defined? Most compilers (for twos-complement architectures at least) are not using that latitude, and I would also guess that most programs that are written for twos-complement arithmetic likewise not expecting undefined behavior for non-overflowing left shifts of negative numbers. Thanks!
"Implementation-defined" is a nuisance, because then you need to add code for all the variations, which also requires a set of standard macros, etc. It is easier and less trouble-prone to just avoid using the currently undefined behavior.
