Remove internal memcpy implementation

[hkratz]

LLVM is unrolling it for the 128+ bytes case which never happens. The speedup with this patch is negligable but code size is reduced quite a bit.

It is a shame that all attempts to do this the Rust way (using copy_from_slice() or read_unaligned()/write_unaligned()) have resulted in non-inlined calls to memcpy with noticable slowdown.

This implementation is optimally auto-vectorized and inlined. The compiler can even prove that the len is less < 32 at the first call site and optimizes that check away.

[lemire]

non-inlined calls to memcpy with noticable slowdown

I find it surprising that call to memcpy would be inefficient. It is not like the function call itself ought to be expensive.

[hkratz]

I find it surprising that call to memcpy would be inefficient. It is not like the function call itself ought to be expensive.

That was a decision I made early on when I was also testing it on inputs < 64 bytes. I will benchmark it again against the memcpy call and throw the custom memcpy out if there is no clear win.