[IR] Maximum & Minimum

[xeren]

Adds signed and unsigned integer maximum and minimum as binary operations. Adds support for respective atomic operations in LLVM IR.

[hernanponcedeleon]

We should also be able to simplify much of the code from #937, right?

[ThomasHaas]

I don't think there is much code on #937 to be simplified, is there? It looks like you would only replace some ITE(LTE(..), .., ..) construct with a single operator but that's it.

Btw. I think a better long-term solution would actually be to have "Expressions with holes" and then instead of generating an RMWOp with some (op, operand)-pair you would just provide the expression. Then we don't even need ops for min/max, because you could put min(_, val) := ITE( _ < val, _, val) as RMW operation where _ gets replaced by the register you load into. Basically, an RMWOp could compute an arbitrary expression over the loaded value instead of the handful of predefined one's.

EDIT: I just prototyped a version for LLVM_rmw_op with holes like so (taken from an IMM test)

8:       bv32 r3 = llvm_rmw_op(&x, bv32 {_} + bv32(1), REL) // amounts to fetch_add(&x, 1, REL)

I think it is certainly feasible and would could unify a lot. It would also cover Xchg (don't add a hole), possibly any form of CAS by adding an optional cmp-expr with hole, and generally also any fancy operation we can think of in our expression system, e.g. atomic_fetch_cttz, and any atomics on floats.