To: J3                                                     J3/26-176
From: Malcolm Cohen
Subject: Some US19 new intrinsics are not useful enough
Date: 2026-June-15


1. Introduction

Some of the new intrinsics inserted by US19 are not very useful in themselves,
and are peculiar variations on ordinary mathematical functions whose only
purpose is to avoid rounding errors that could easily be avoided by the usual
mathematical functions.

It is unhelpful, in my opinion, to put the burden of accuracy onto the user
writing the program. Especially since as a recommendation the hoped-for
reduction in rounding errors might not materialise.

We should reconsider the approach.


2. Compound interest

This function is appropriate for COBOL, but hardly for Fortran.
Furthermore, writing 1.05**N for 5% compound interest over N years is trivial.
Having this as a separate function is a waste of time.

I note that IEEE require correct rounding. That is, in my opinion, pointless,
as in real life, compound interest is rounded each year, not after N years.


3. LOG and EXP with plus/minus one

Instead of a bunch of new functions, we could simply recommend that
    "For LOG(<one>+expr) or LOG(expr+<one>), where <one> is a constant
     expression equal to one, the processor should compute the result without
     losing precision by adding <one> to the constant expression."
and
    "In the expression EXP(x)-<one>, where <one> is a constant expression
     equal to one, the processor should compute the value without losing
     precision by forming EXP(x) and then subtracting one."

This avoids distorting the formulae simply to accommodate computations with
values very close to one.

A NOTE would also be appropriate, viz
   "NOTE
    The recommendation means that it would be expected that LOG(1+<x>) has the
    same value as the C function logp1(<x>)."

===END===