CIRGen support for `*_atomic_{thread,signal}_fence`

[bcardosolopes]

All the below are missing:

CodeGen/CIRGenBuiltin.cpp:  case Builtin::BI__atomic_thread_fence:
CodeGen/CIRGenBuiltin.cpp-  case Builtin::BI__atomic_signal_fence:
CodeGen/CIRGenBuiltin.cpp:  case Builtin::BI__c11_atomic_thread_fence:
CodeGen/CIRGenBuiltin.cpp-  case Builtin::BI__c11_atomic_signal_fence:
CodeGen/CIRGenBuiltin.cpp:    llvm_unreachable("BI__atomic_thread_fence like NYI");

[elhewaty]

I will give it a try, assign me please

[bcardosolopes]

@elhewaty awesome, just did!

[Rajveer100]

@bcardosolopes
Could you provide some context here with respect to the implementations?

From my understanding, this is similar to adding an operation in MLIR dialect, and then referencing with the cir:: namespace.

[bcardosolopes]

@Rajveer100 sure, we probably want a new cir::FenceOp that encodes (a) the sync scope somehow (perhaps a attribute enum kind that covers single_thread and system) and (b) the atomic ordering. If the atomic ordering isn't constant, we don't want to early expand it during CIRGen, but only during LLVM Lowering. Lowering to LLVM should also be implemented as part of the PR that fixes this.