[Mlir-commits] [mlir] [mlir][llvmir] Add new support for strict fp handling (PR #205158)

[llvmlistbot at llvm.org]

https://github.com/adams381 requested changes to this pull request.

I think this is the right approach, and lowering to the existing constrained intrinsics for now, so the IR representation can change later without disturbing the attribute, is what makes me comfortable landing it ahead of Nikita's IR RFC.  The conditional-purity modeling and the `ConstrainedOps.def`-based mapping both look right to me.

The rounding-mode lowering looks wrong, though; I've put the details inline and I think it's a correctness issue rather than a preference.  There's also a semantics question on the attribute builder and a couple of smaller things.

Not for this PR, but worth keeping in mind: once the arith and cir copies land, the three attributes only stay interchangeable if their enums and field semantics match exactly, so sharing the enum definitions rather than retyping them seems worth it.

https://github.com/llvm/llvm-project/pull/205158