[PATCH] D70261: [strictfp] Add token support for FP constraints

[Andy Kaylor via Phabricator via llvm-commits]

andrew.w.kaylor added a comment.

As referenced in Simon's reply, I have started a mailing list discussion about this.

I see your point, Ulrich, about this overloading the token concept a bit. The token type is also used by coroutine handling, but it uses tokens in the same way as exception handling -- it's an opaque value returned by an intrinsic call and passed to other intrinsics. In what I've proposed here, the optimizer is intended to inspect and use the value of the token. However, it's still opaque to the program. For instance, you can't assign the token to another token or add two tokens together or print the token to stdout.

I feel pretty good about moving the constraints to an operand bundle and making it optional. I'm on the fence about tokens as the way to do it.

I don't think any current concept fits better (maybe attributes, but call site specific attributes are uncommon and confusing). A new IR concept to handle this feels like a worse idea than extending the concept of tokens.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D70261/new/

https://reviews.llvm.org/D70261