Avoid link-time dependency on core through to_bits and from_bits

[tmiasko]

The compiler-builtins should not have any link time dependencies on any
other crates including core.

Replace uses of to_bits and from_bits from core with local repr
and to_repr methods implemented in terms of the transmute intrinsic.

This fixes a build issue that can occur in build configurations that
share generic monomorphizations between crates.

[bjorn3]

These functions should always be inlined, right? They are marked as #[inline] and small enough that no reasonable compiler would not inline them when optimizations are enabled.

[tmiasko]

Those functions themselves are indeed marked inline but not everything they transitively use is.

[bjorn3]

What do you mean? f*::to_bits and f*::from_bits are both marked as #[inline] and only call mem::transmute which is an intrinsic and thus by definition inlined.

[tmiasko]

to_bits calls const_eval_select, which calls call_once, which calls a closure, which calls the intrinsic.

[bjorn3]

I see. On stable it is a simple transmute. On nightly it uses const_eval_select. Making the rt_f32_to_u32 closure a regular function and marking it as #[inline] should help with that I think.

[tmiasko]

Given the role of compiler builtins, I would assume that limiting the dependency on the core is desirable regardless, so this was my first choice in terms of the implementation strategy. Modifying the core instead would presumably work as well.

[Amanieu]

We can't realistically avoid all dependencies on core, we just have to make sure that we only use inline functions. It is concerning however that to_bits/from_bits is not properly inlined in core and this should be fixed there.

[tmiasko]

I opened a pull request to modify the core instead. Though, that approach is not as robust as solution proposed here since it requires enabled optimizations when buliding the core.