Fix resizeRawFloat preserving the correct sExp for NaN/Infty inputs

[jerryz123]

The sAdjustedSexp calculation does not correctly represent the sExp field when the input is NaN or infinity. This corrects that, I believe.

[aswaterman]

Curious where this is cropping up.

The HardFloat spec says that this should not be needed, as the other fields in the RawFloat are don't-cares if isInf or isNaN is true (excepting NaN payload propagation, which doesn't matter for us). See Sec. 5.3 of https://www.jhauser.us/arithmetic/HardFloat-1/doc/HardFloat-Verilog.html

[jerryz123]

I'm using resizeRawFloat to upconvert sub-F64 values such that I can use the same F64 RecFNToIN to do IEEE fp-to-int conversions.

The problem I'm seeing is that the conversion of a rawFN to a recFN expressed here:

berkeley-hardfloat/hardfloat/src/main/scala/recFNFromFN.scala

Lines 47 to 51 in 26f00d0

	rawIn.sign ##
	(Mux(rawIn.isZero, 0.U(3.W), rawIn.sExp(expWidth, expWidth - 2)) |
	Mux(rawIn.isNaN, 1.U, 0.U)) ##
	rawIn.sExp(expWidth - 3, 0) ##
	rawIn.sig(sigWidth - 2, 0)

does not construct a correct recFN representation of a infinity or NaN when the input rawFloat is resized up (the high non-sign bits of the exponent are not 11).

Perhaps this expression is not intended to work with resized rawFNs?

[aswaterman]

Yeah, that expression is only valid in its context; it is (legitimately) making assumptions about what rawFloatFromFN might return.