Implementation restriction: although numeric values have arbitrary precision in the language, implementations may implement them using an internal representation with limited precision. That said, every implementation must:

Represent integer values with at least 256 bits.

Represent floating-point values with a mantissa of at least 256 bits and a signed binary exponent of at least 16 bits.

Give an error if unable to represent an integer value precisely.

Give an error if unable to represent a floating-point value due to overflow.

Round to the nearest representable value if unable to represent a floating-point value due to limits on precision. These requirements apply to the result of any expression except for builtin functions, for which an unusual loss of precision must be explicitly documented.

[–] deegeese@sopuli.xyz 4 points 1 month ago* (last edited 1 month ago)

Thanks for teaching me something, but the obscurity of your answer just illustrates how rare that requirement is in human readable formats, and mostly limited to data formats designed for numeric precision, like HDF5, FITS or protobuf.