Keep higher level abstraction for runtime memory orders on atomics

[bcardosolopes]

#731 adds CIRGen support for non-constant memory orders all the way down to LLVM.

To keep CIR a bit more simple and easier to analyze, we could change atomic operations to support the non-constant version as well and do the same lowering during LoweringPrepare instead (see clang/test/CIR/CodeGen/atomic-runtime.cpp for the example currently generated). This also opens the possibility of folding atomic operations into the constant version once we start doing some inline in CIR.

[bcardosolopes]

Sorry about the delay!

However, is there another ideal solution that I might have overlooked?

I think it might be better just to create a new operation, more clean and direct. You can refactor AtomicCmpXchg into a tablegen class (see how we implement CallOp) if you want to reuse part of the code.

If we expand the non-constant cir.atomic.cmp_xchg operation to a constant version during LoweringPrepare,

This should be done in two steps:

1. CIRSimplify transforms the dynamic version into the constant one if possible
2. If it gets to LoweringPrepare, it means it wasn't able to optimize and should just lower.

we may need to clone the cir.unary, cir.if, and cir.store operations into switchop

Not sure I follow, if you are able to transform it to the constant version, why do you need a switch anyways?