Stage 2 llvm backend, improve code gen for return values

[lerno]

Currently the ABI for passing optional and error parameters prevent full optimization by LLVM.

1. Sub optimal code generation with full optimization when using catch |...|
   https://godbolt.org/z/xqTKWKW37

2. Optimal code generation with full optimization when performing essentially the same but using catch 1
   https://godbolt.org/z/PWe4GPb3c

For Optional, both variants produce sub optimal code:
https://godbolt.org/z/qGavK9aWK

Note that Swift has a similar strategy (passing a struct), but manages to get better optimization:
https://godbolt.org/z/4xa5Y85Gx

[lerno]

It is possible to follow the Swift implementation to get good optimization.

Another method would be to return the error as the normal return value and have the normal return out to be passed by ref. E.g. fn square(n:i32) !i32 is lowered to fn square(...) i16

[lerno]

A related, possible enhancement which will allow dynamic linking is the following way to create an error code:

For every error code:

1. Create a unique global
2. Use the (constant) address to this global as the value of the error code.

The (dynamic) C linker will ensure that all libraries use the same address, which ensures uniqueness. This is also understood by C code. This would make the error code pointer sized, but if the error is passed as a return value, then this would usually be exactly register sized and consequently equally cheap as using a i16.

[andrewrk]

Related: #561