CArray.make with initial value is extremely slow

[Lupus]

Looks like initialization of each byte goes through C stub call. Initialization of 40MB buffer took about a second on a server with modern intel CPU. Can this be achieved by calling memset? At least zero-filling probably needs some optimized specialization...

[yallop]

Indeed, this is a naive implementation, and it should be possible to do considerably better. Several possible improvements, some more work than others:

• Eliminate allocations from the loop (which currently creates a new fat pointer for each element in unsafe_set)
• Eliminate dispatch/branching from the loop (which currently inspects the type representation once per element in Ctypes_memory.stubs and then again in ctypes_write)
• Double-check that the ctypes_write function doesn't allocate, and mark it noalloc if so
• Eliminate the C-call-per-element by moving the loop into C
• (Perhaps) add specialized implementations for commonly-used types. For char we could use memset

At least zero-filling probably needs some optimized specialization...

That would be useful. As it happens, ctypes uses calloc to allocate C arrays, so they're already guaranteed to be zero-filled when no initializer is supplied.