I think secp256k1 needs an explicit no-floating point policy, like the Linux kernel. This means no use of float, and no double (and no half, no long double either).

• It has to be able to run in kernel context
• Boot and security processors tend to have no FP support (e.g. core 0 on the Sifive U54)
• FP hardware implementation is possibly inconsistent between platforms which is useless for cryptography

I mean for the main code; in the bench framework is fine, though with the assumption that it might be emulated and very slow.

(inspired by #658 (comment))