New issue from Jan Schultke: ""round_to_nearest" rounding mode is unclear"

Author: Dani-Hub

xml/issue4474.xml

@@ -0,0 +1,79 @@
+<?xml version='1.0' encoding='utf-8' standalone='no'?>
+<!DOCTYPE issue SYSTEM "lwg-issue.dtd">
+
+<issue num="4474" status="New">
+<title>"`round_to_nearest`" rounding mode is unclear</title>
+<section><sref ref="[round.style]"/></section>
+<submitter>Jan Schultke</submitter>
+<date>13 Nov 2025</date>
+<priority>99</priority>
+
+<discussion>
+<p>
+Consider the specification of `round_to_nearest` in <sref ref="[round.style]"/>:
+</p>
+<blockquote><p>
+`round_to_nearest` if the rounding style is to the nearest representable value
+</p></blockquote>
+<p>
+It is unclear how exact ties are rounded. For example, with this rounding, would a value
+that is equidistant to zero and <tt>numeric_limits&lt;float&gt;::min()</tt> be rounded 
+towards zero or away from zero?
+<p/>
+In <sref ref="[numeric.limits.members]"/>, there exists a footnote for <tt>numeric_limits&lt;T&gt;::round_style</tt>:
+</p>
+<blockquote><p>
+185) Equivalent to `FLT_ROUNDS`. Required by ISO/IEC 10967-1:2012.
+</p></blockquote>
+<p>
+In C23 5.2.4.2.2 [Characteristics of floating types <tt>&lt;float.h&gt;</tt>], it is specified 
+that a value of `1` for `FLT_ROUNDS` (which equals `round_to_nearest`) means
+</p>
+<blockquote><p>
+to nearest, ties to even
+</p></blockquote>
+<p>
+This is also the default ISO/IEC 60559 rounding mode, and chosen by standard libraries such as 
+libstdc++ for `numeric_limits` under that assumption. Do note that C23 no longer references 
+ISO/IEC 10967 in any normative wording, so presumably, matching `FLT_ROUNDS` values means to match 
+the value that exists in C23, including its meaning.
+</p>
+</discussion>
+
+<resolution>
+<p>
+This wording is relative to <paper num="N5014"/>.
+</p>
+
+<ol>
+<li><p>Modify <sref ref="[round.style]"/> as indicated:</p>
+
+<blockquote>
+<p>
+-1- The rounding mode for floating-point arithmetic is characterized by the values:
+</p>
+<ul style="list-style-type: none">
+<li>(1.1) &mdash; 
+`round_indeterminate` if the rounding style is indeterminable
+</li>
+<li>(1.2) &mdash; 
+`round_toward_zero` if the rounding style is toward zero
+</li>
+<li>(1.3) &mdash; 
+`round_to_nearest` if the rounding style is to the nearest representable value<ins>; if there are 
+two equally near values, the one with an even least significant digit is chosen</ins>
+</li>
+<li>(1.4) &mdash; 
+`round_toward_infinity` if the rounding style is toward infinity
+</li>
+<li>(1.5) &mdash; 
+`round_toward_neg_infinity` if the rounding style is toward negative infinity
+</li>
+</ul>
+
+</blockquote>
+</li>
+</ol>
+</resolution>
+
+</issue>