Hexadecimal float literal rounding spec vs test

[AndrewScheidecker]

This is asserting that a hexadecimal float literal with too many significand bits is rejected as malformed.

I believe this doesn't match the spec, which implies that a hexadecimal float literal with too many hexits may be accepted with rounding:

The value of a literal must not lie outside the representable range of the corresponding IEEE 754 type (that is, a numeric value must not overflow to ±infinity), but it may be rounded to the nearest representable value.

Note: Rounding can be prevented by using hexadecimal notation with no more significant bits than supported by the required type.

Is the intent for the spec to prohibit rounded hex floats?

[rossberg]

Extra digits are fine, and will be rounded-to-nearest ties-to-even, but it is an overflow if this causes rounding to infinity -- which is the case for the particular example you link to (f32 has a 23 bit mantissa, so can represent up to 0x1.fffffep127 exactly; the value 0x1.fffffe7p127 still rounds down to that, but 0x1.ffffffp127 rounds up to infinity).

[AndrewScheidecker]

Thanks, that makes sense.