I’d like to propose starting a working group to discuss various issues related to floating-point handling in LLVM IR. Following up on the discussion in [FP] Constant folding of floating-point operations I’ve started a document which I hope will become a normative reference for floating-point semantics in LLVM IR, but the nature of this document requires consensus so rather than just put down what I want and try to push it through, I thought it would be good if we could create a working group to hammer out all the details.
I would certainly be interested in joining such a working group.
Tom Honermann’s recent announcement that he’s working on a decimal floating-point implementation brings to mind that that should also be added to the topics for discussion, especially because decimal floating-point adds issues not present or less relevant for binary floating-point (preferred exponent and noncanonical values, in particular).
It occurs to me that I should probably ask interested people to indicate their time zone. I imagine any time will be inconvenient for some people, but I’ll try not to make it egregiously so. I’m in the Pacific time zone.
Apologies to anyone who was watching this and waiting for me to start the proposed group. I’ve been busy with other tasks, but I hope to get back to this very soon.
I’ve just added something to the LLVM Calendar for the third Wednesday of every month, starting November 15 at 1 PM PST/9 PM UTC. We can adjust that as needed based on who wants to participate and when they’re available.