Fewer rounding errors with (format "%f" fixnum)

* etc/NEWS: Document this.
* src/editfns.c (styled_format): When formatting integers via a
floating-point format, use long double instead of double
conversion, if long double’s extra precision might help.

1 files changed, 8 insertions, 0 deletions

diff --git a/etc/NEWS b/etc/NEWS
index a8db54c51ef..9c995930fbf 100644
--- a/etc/NEWS
+++ b/etc/NEWS
@@ -907,6 +907,14 @@ compares their numerical values.  According to this predicate,
 due to internal rounding errors.  For example, (< most-positive-fixnum
 (+ 1.0 most-positive-fixnum)) now correctly returns t on 64-bit hosts.
 
+---
+** On hosts like GNU/Linux x86-64 where a 'long double' fraction
+contains at least EMACS_INT_WIDTH - 3 bits, 'format' no longer returns
+incorrect answers due to internal rounding errors when formatting
+Emacs integers with %e, %f, or %g conversions.  For example, on these
+hosts (eql N (string-to-number (format "%.0f" N))) now returns t for
+all Emacs integers N.
+
 +++
 ** The new function 'char-from-name' converts a Unicode name string
 to the corresponding character code.