implementations, I stumbled on some issues (those are issues in test code, not the spec).

* Some test data is desired to have more precision, which I created a PR.

  https://github.com/scheme-requests-for-implementation/srfi-144/pull/7

* A couple of tests compared with exact match may yield 1ulp-off result computed

  in 64bit IEEE double precision throughout.  (It might not show if intermediate results

  is calculated in higher precision)

 -   (* 2 fl-1/sqrt-2)

 -   (fl* (flonum 1/3) (flonum 1/3) (flonum 1/3)), compared to (flonum (expt 1/3 3))

- Definition of flnormalized-fraction-exponent in the spec doesn't mention what if

  the argument is infinity.  The spec mentions C99 frexp, which returns +/-inf  as the

  fraction if the argument is +/-inf.  The reference implementation returns 0.5 as the

  normalized fraction, which does fits the description (0.5 <= abs(fraction)  < 1.0)

  but it also returns a finite exponent value which doesn't fit the description

  (although it is larger than the possible value).  So at best I read the behavior is

  undefined in spec.  The tests assume 0.5 is returned in such case.