Revisit NOP cases for intrinsics

[RalfJung]

copy, copy_nonoverlapping and write_bytes currently accept bad pointers when the size is 0, because libstd relies on this.

I think we should enforce the check, and fix libstd.

[eddyb]

This is a point of debate in C, too, isn't it? Do we have something conclusive to rely on, and enforce?

[RalfJung]

It's UB in C, but compilers do not dare exploiting that in general.

I wrote to llvm-dev last year and was unable to get a definite answer out of them, see this mailing list thread.

So, the safest thing to do is to say that it's UB for us...

[RalfJung]

I just opened a bug on their side to maybe get something moving: https://bugs.llvm.org/show_bug.cgi?id=38583