To: J3                                                     J3/26-211
From: Malcolm Cohen
Subject: Interp F23/027 on IEEE_SCALB and non-integral scaling
Date: 2026-October-20


----------------------------------------------------------------------

NUMBER: F23/027
TITLE: IEEE_SCALB is not defined for finite non-integral values of I
KEYWORDS: IEEE_SCALB, ISO/IEC/IEEE 60559
DEFECT TYPE: Erratum
STATUS: J3 consideration in progress
REFERENCES: 24-007, N2244 (F2023 draft corrigendum 1), N2258

QUESTION:

The description of IEEE_SCALB does not describe the result when I is
finite and non-integral.

ISO/IEC/IEEE 60559 says that says non-integral I is language-defined,
but we do not define it.

That means that a program which does
    PRINT *,IEEE_SCALB(x,1.54)
has no established interpretation, and is thus non-conforming.
That is probably an overreaction to a numerical error.

The C standard does not have a floating-point I argument,
so we cannot get guidance from them.

There are two obvious possibilities:
    (1) raise INVALID, return NaN, because you cannot add a noninteger
        value to the exponent.
    (2) raise INEXACT and return X*(RADIX(X)**I). That would require a
        significant effort if we want it to avoid overflow/underflow
        in the calculation of RADIX(X)**I. It also fails to alert the
        user to the mistake (no-one looks at INEXACT, which was likely
        already set from an earlier calculation anyway).

What should the result be, or should this continue to be
completely non-conforming.

ANSWER:

This was not intended to be non-conforming.
The robust result is that INVALID should be raised,
and the result should be a quiet NaN.

An edit is provided to correct this oversight.

EDIT to 24-007.

[488:16+] 17.11.37 IEEE_SCALB, p6 Result Value,
          Append new case
   "Case (v): If X and I are finite, and I is non-integral, the
              IEEE_INVALID exception shall occur. If
              IEEE_SUPPORT_NAN(X) is true, the result value is a
              quiet NaN."

EDIT to N2258.

[581:13+] 18.11.37 IEEE_SCALB, p6 Result Value,
          Append new case
   "Case (v): If X and I are finite, and I is non-integral, the
              IEEE_INVALID exception shall occur. If
              IEEE_SUPPORT_NAN(X) is true, the result value is a
              quiet NaN."

SUBMITTED BY: Malcolm Cohen

HISTORY: 26-nnn   m241  Submitted

----------------------------------------------------------------------