bug#31938: regression in (format "%d" -0.0)

bug#31938: regression in (format "%d" -0.0)

[Paul Pogonyshev]

Severity: serious

Emacs 24-26:

    (format "%d" -0.0) => "0"

Emacs 27:

    (format "%d" -0.0) => "-0"

I think this is a very important regression, because it happens in
very low-level code and can lead to unpredictable results in certain
special cases. Caught with real-world `datetime' package: on Emacs 27
all its regression tests fail because of this change, on Emacs 24-26
they pass.

In the library, -0.0 comes from the fact that `mod' built-in can
return it as a result (it is probably fine, because it is _equal_ to
0.0, so it is not negative, even if it looks like it).

I.e. real code looks more like this:

    (let ((x -2.0)) (format "%d" (mod x 2)))

which gives "-0" and is a regression in Emacs 27.

bug#31938: followup

[Paul Pogonyshev]

Another difference. In Emacs 24-26:

    (format "%d" 0.9) => "0"

In Emacs 27:

    (format "%d" 0.9) => "1"

bug#31938: followup

[Eli Zaretskii]

> From: Paul Pogonyshev <pogonyshev@gmail.com>
> Date: Sat, 23 Jun 2018 12:21:00 +0200
> 
> Another difference. In Emacs 24-26:
> 
>     (format "%d" 0.9) => "0"
> 
> In Emacs 27:
> 
>     (format "%d" 0.9) => "1"

This looks like the (unintended?) result of 80e145fc9.  CC'ing Paul
who made that change.

(FWIW, I'm not sure why this is a problem.)

P.S. Please don't change the Subject when you post to a bug, keep the
original Subject you used when reporting it.  That makes looking for
related messages in a mailer easier, at least for me.

bug#31938: followup

[Paul Eggert]

[-- Attachment #1: Type: text/plain, Size: 1197 bytes --]

On 06/23/2018 04:46 AM, Eli Zaretskii wrote:
>> From: Paul Pogonyshev <pogonyshev@gmail.com>
>> Date: Sat, 23 Jun 2018 12:21:00 +0200
>>
>> Another difference. In Emacs 24-26:
>>
>>      (format "%d" 0.9) => "0"
>>
>> In Emacs 27:
>>
>>      (format "%d" 0.9) => "1"
> This looks like the (unintended?) result of 80e145fc9.

I vaguely recall relying on the Elisp manual, which says that %d must be 
applied only to integers, and I guess I figured the semantics didn't 
matter so that the simplest implementation (thought that rounding and 
keeping -0) would satisfy the doc and anyway would be more helpful than 
truncating and discarding the sign of -0. But now that I reread the 
manual I see that it says %d must signal an error when given a float, 
which is obviously not the behavior, so the manual is wrong. And now 
that this issue has come up I notice that 80e145fc9 caused %d to behave 
inconsistently with %x; the latter truncates and converts -0 to 0 
whereas the former rounds and keeps -0. This inconsistency is not good.

To try to improve this, I installed the attached into master to change 
'format' to be more compatible with Emacs 26 etc. and am boldly closing 
the bug report.

[-- Attachment #2: 0001-format-d-F-now-truncates-floating-F.patch --]
[-- Type: text/x-patch, Size: 5175 bytes --]

From 32a9df02ac3c5b12d7ff6efa0c3924d740d70268 Mon Sep 17 00:00:00 2001
From: Paul Eggert <eggert@cs.ucla.edu>
Date: Mon, 25 Jun 2018 12:21:40 -0700
Subject: [PATCH] (format "%d" F) now truncates floating F

Problem reported by Paul Pogonyshev (Bug#31938).
* src/editfns.c: Include math.h, for trunc.
(styled_format): For %d, truncate floating-point numbers and
convert -0 to 0, going back to how Emacs 26 did things.
* doc/lispref/strings.texi (Formatting Strings):
Document behavior of %o, %d, %x, %X on floating-point numbers.
* src/floatfns.c (trunc) [!HAVE_TRUNC]: Rename from emacs_trunc
and make it an extern function, so that editfns.c can use it.
All callers changed.
* test/src/editfns-tests.el (format-%d-float): New test.
---
 doc/lispref/strings.texi  | 11 ++++++++---
 src/editfns.c             |  7 +++++++
 src/floatfns.c            | 13 +++++--------
 src/lisp.h                |  5 ++++-
 test/src/editfns-tests.el |  8 ++++++++
 5 files changed, 32 insertions(+), 12 deletions(-)

diff --git a/doc/lispref/strings.texi b/doc/lispref/strings.texi
index 70ba1aa613..026ba749cb 100644
--- a/doc/lispref/strings.texi
+++ b/doc/lispref/strings.texi
@@ -922,18 +922,23 @@ Formatting Strings
 @item %o
 @cindex integer to octal
 Replace the specification with the base-eight representation of an
-unsigned integer.
+unsigned integer.  The object can also be a nonnegative floating-point
+number that is formatted as an integer, dropping any fraction, if the
+integer does not exceed machine limits.
 
 @item %d
 Replace the specification with the base-ten representation of a signed
-integer.
+integer.  The object can also be a floating-point number that is
+formatted as an integer, dropping any fraction.
 
 @item %x
 @itemx %X
 @cindex integer to hexadecimal
 Replace the specification with the base-sixteen representation of an
 unsigned integer.  @samp{%x} uses lower case and @samp{%X} uses upper
-case.
+case.  The object can also be a nonnegative floating-point number that
+is formatted as an integer, dropping any fraction, if the integer does
+not exceed machine limits.
 
 @item %c
 Replace the specification with the character which is the value given.
diff --git a/src/editfns.c b/src/editfns.c
index 30d585cd01..7d032a7ca4 100644
--- a/src/editfns.c
+++ b/src/editfns.c
@@ -47,6 +47,7 @@ along with GNU Emacs.  If not, see <https://www.gnu.org/licenses/>.  */
 #include <errno.h>
 #include <float.h>
 #include <limits.h>
+#include <math.h>
 
 #ifdef HAVE_TIMEZONE_T
 # include <sys/param.h>
@@ -4671,6 +4672,12 @@ styled_format (ptrdiff_t nargs, Lisp_Object *args, bool message)
 		    {
 		      strcpy (f - pMlen - 1, "f");
 		      double x = XFLOAT_DATA (arg);
+
+		      /* Truncate and then convert -0 to 0, to be more
+			 consistent with %x etc.; see Bug#31938.  */
+		      x = trunc (x);
+		      x = x ? x : 0;
+
 		      sprintf_bytes = sprintf (sprintf_buf, convspec, 0, x);
 		      char c0 = sprintf_buf[0];
 		      bool signedp = ! ('0' <= c0 && c0 <= '9');
diff --git a/src/floatfns.c b/src/floatfns.c
index ec0349fbf4..e7d404a84e 100644
--- a/src/floatfns.c
+++ b/src/floatfns.c
@@ -435,11 +435,9 @@ emacs_rint (double d)
 }
 #endif
 
-#ifdef HAVE_TRUNC
-#define emacs_trunc trunc
-#else
-static double
-emacs_trunc (double d)
+#ifndef HAVE_TRUNC
+double
+trunc (double d)
 {
   return (d < 0 ? ceil : floor) (d);
 }
@@ -482,8 +480,7 @@ Rounds ARG toward zero.
 With optional DIVISOR, truncate ARG/DIVISOR.  */)
   (Lisp_Object arg, Lisp_Object divisor)
 {
-  return rounding_driver (arg, divisor, emacs_trunc, truncate2,
-			  "truncate");
+  return rounding_driver (arg, divisor, trunc, truncate2, "truncate");
 }
 
 
@@ -543,7 +540,7 @@ DEFUN ("ftruncate", Fftruncate, Sftruncate, 1, 1, 0,
 {
   CHECK_FLOAT (arg);
   double d = XFLOAT_DATA (arg);
-  d = emacs_trunc (d);
+  d = trunc (d);
   return make_float (d);
 }
 \f
diff --git a/src/lisp.h b/src/lisp.h
index d0c52d8567..8c884dce15 100644
--- a/src/lisp.h
+++ b/src/lisp.h
@@ -3425,8 +3425,11 @@ extern Lisp_Object string_make_unibyte (Lisp_Object);
 extern void syms_of_fns (void);
 
 /* Defined in floatfns.c.  */
-extern void syms_of_floatfns (void);
+#ifndef HAVE_TRUNC
+extern double trunc (double);
+#endif
 extern Lisp_Object fmod_float (Lisp_Object x, Lisp_Object y);
+extern void syms_of_floatfns (void);
 
 /* Defined in fringe.c.  */
 extern void syms_of_fringe (void);
diff --git a/test/src/editfns-tests.el b/test/src/editfns-tests.el
index 1ed0bd5bba..c828000bb4 100644
--- a/test/src/editfns-tests.el
+++ b/test/src/editfns-tests.el
@@ -176,6 +176,14 @@ transpose-test-get-byte-positions
   (should-error (format "%o" -1e-37)
                 :type 'overflow-error))
 
+;; Bug#31938
+(ert-deftest format-%d-float ()
+  (should (string-equal (format "%d" -1.1) "-1"))
+  (should (string-equal (format "%d" -0.9) "0"))
+  (should (string-equal (format "%d" -0.0) "0"))
+  (should (string-equal (format "%d" 0.0) "0"))
+  (should (string-equal (format "%d" 0.9) "0"))
+  (should (string-equal (format "%d" 1.1) "1")))
 
 ;;; Check format-time-string with various TZ settings.
 ;;; Use only POSIX-compatible TZ values, since the tests should work
-- 
2.17.1