> However, when v has an arbitrary type, it will instead just clone() it for every single element of the array, which is many orders of magnitude slower. So much for zero cost abstraction.
Arguably the WrapperType, i.e. any arbitrary type is the general case and thus the baseline performance. `u8` is the special case (hence specialization) that performs the alloc_zeroed optimization. So it's not really that the abstraction adds a cost. It's just that the special case removes a cost paid by everything else.
In the future the vec initialization might be fixable, but this requires turning potentially undefined values into some valid if arbitrary bit patterns (i.e. llvm's freeze) to compare it against zero.
Arguably the WrapperType, i.e. any arbitrary type is the general case and thus the baseline performance. `u8` is the special case (hence specialization) that performs the alloc_zeroed optimization. So it's not really that the abstraction adds a cost. It's just that the special case removes a cost paid by everything else.
In the future the vec initialization might be fixable, but this requires turning potentially undefined values into some valid if arbitrary bit patterns (i.e. llvm's freeze) to compare it against zero.